AI Workflows in the wallet¶
Release status
AI Workflows ships since wallet 0.8.8 (test-network pre-release, CR-0044). The nodes and templates marked new in 0.8.9 below were added in wallet 0.8.9 (CR-0052).
Step-by-step user guide: pqwerai.com/wallet/workflows/.
The AI Workflows page (Ctrl+Shift+8) lets you build an AI task as connected nodes, then run it. Every node calls a real wallet service. Nothing is simulated: a step either runs or stops with the reason and a suggested fix.
The Canvas page (Ctrl+Shift+7) is a separate tool. It builds Solidity contracts and is unchanged.
Simple and Advanced modes¶
| Mode | What you see |
|---|---|
| Simple | Templates, the template's text inputs, a Run button and the results |
| Advanced | The node library and search, every parameter, groups, subworkflows, the problems list and the run log |
Starter templates¶
Every template was run live on 19 November 2026 the way a new user meets it: the approved local model, the public
chain-20777 RPC and AI market, an empty data folder and every payment refused (test
WorkflowTemplateAuditLiveTests, CR-0076). Each template now says under its name what it needs; templates that read a
file copy a sample into the workflow data folder when you open them, never over a file you already have.
| Template | What it does | Cost | Live result for a new user |
|---|---|---|---|
| AI chat | Asks your local model; a Memory node keeps the conversation | Free (local) | Works |
| Multi-agent collaboration | A planner, a critic and a writer take turns on one task | Free (local) | Works |
| MeshLLM paid inference | Finds providers on chain, picks the cheapest, shows the price, pays through escrow | Paid: asks first | Finds the provider and the signed quote, then stops at your approval |
| AI training: evaluate a model | Scores your local model on a JSONL dataset | Free (local) | Works with the sample qa.jsonl |
| Smart-contract interaction | Reads a contract on chain 20777, checks the result and pins it to IPFS | Free | Needs your IPFS node running; the contract read works without it |
| Summarize and translate | Bullet-point summary with the local model, then a translation | Free (local) | Works on a sample text (before CR-0076 it summarized its own placeholder) |
| Sort messages by type | Classify labels a message; each label goes to its own branch | Free (local) | Works |
| Extract data to JSON | Extract Fields fills named fields, Get Field and Check verify the total, File saves the JSON | Free (local) | Works |
| Balance report | Block Height and Balance feed a Template that is shown and saved to a file | Free | Works (reads the IPFS reward pool's balance) |
| Ask your documents (RAG) | Splits and indexes documents, answers with citations | Free (local) | Works with the sample handbook.md (keyword search; connect an embedding model for semantic search) |
| LoRA style | Merges LoRA adapters into a base model and chats with it | Free (local) | Needs your own base model and LoRA files; the built-in Bonsai model cannot take a LoRA |
Before you run, reused steps and run history (CR-0077)¶
These features are in the development build. They are not in a released wallet yet.
- Before you run. The Run tab lists what the workflow needs, line by line, with the node and the fix:
- the files in the data folder,
- whether the local AI model is running,
- whether your IPFS node is running,
- the network connection,
- the most the run can spend.
Check runs these tests without running any step. Run runs them first, then runs the workflow anyway: a step that misses something fails with the same fix. Nothing is signed or paid by the check. - Reuse unchanged results. A local step whose settings and inputs did not change since the last run is not run again. It shows Cached, so a long model answer is not recomputed when you only change a later step. These steps always run again: - steps that read the network or the chain, - steps that write files or use conversation memory, - steps that sign or pay, - agents with tools, - LoRA merges and model servers.
Run fresh forgets the reused results. They are kept in memory for the wallet session only.
- Output of each step. After a run, select a node (Node tab) to see what it produced on each output port, with
its type.
- History. Each run is saved in the workflow folder (runs/<workflow>/, the last 30 per workflow). The record
holds the state, message and output previews of every step. Pick a past run in the History tab to show it on the
canvas. Outputs are previews of at most 4,000 characters. Node outputs never contain keys or passwords.
Models tab, images, video and speech (CR-0079)¶
These features are in the development build. They are not in a released wallet yet.
The Models tab lists the programs and models the canvas can run on your PC. For each one it shows the licence, the download size, whether it fits your GPU, and Install or Remove. Before downloading, the wallet shows every file and the site it comes from, and asks you first. Every byte is checked against a SHA-256 pinned in the wallet. A download that stops resumes where it stopped. Only permissive licences are offered (owner decision, 19 November 2026).
| Entry | What it adds | Licence | Download |
|---|---|---|---|
| Qwen3 0.6B (Q8_0) + TL;DR summary LoRA | A small chat model that can take LoRAs (Bonsai cannot), with a sample LoRA | Apache-2.0 | 650 MB |
| stable-diffusion.cpp (Vulkan) | The image and video program (release master-1106-532d2c1) | MIT | 30 MB |
| Stable Diffusion XL 1.0 + fp16 VAE | 1024×1024 images; needs a GPU with 6 GB or more | CreativeML OpenRAIL++-M / MIT | 7.3 GB |
| Pixel art LoRA for SDXL | A style LoRA for SDXL | CreativeML OpenRAIL-M | 170 MB |
| Wan 2.2 TI2V 5B (Q4_K_M) | Short video clips; needs a GPU with 8 GB or more | Apache-2.0 | 8.5 GB |
| whisper.cpp v1.9.2 + Whisper base | Speech to text in about 99 languages, on the CPU | MIT | 150 MB |
| FFmpeg 8.1 (LGPL build) | Converts recordings for Whisper; runs as a separate program | LGPL | 80 MB |
New nodes (category Image, video & audio):
| Node | What it does |
|---|---|
| Generate Image | Text to image, or image to image with a start image. Uses an installed model plus any LoRA Loaders connected to LoRAs. The same seed and the same inputs give the same image. |
| Generate Video | Text (or a start image) to a short clip with Wan 2.2 5B. |
| Load Image / Load Audio | A file from the workflow data folder. |
| Speech to Text | Whisper transcription. auto detects the language; it can also translate to English. |
| Image / Video (Output) | Shows the result. Open opens the file with your default app. |
LoRAs work the way they do in ComfyUI: chain LoRA Loader nodes with a strength. For text models the stack is merged into the base model. Image LoRAs are applied when the image is made. A LoRA only fits the model family it was trained for. The Models tab names each LoRA's base model, and the check before running refuses a mismatch (for example, a Qwen3 text LoRA on SDXL).
New templates: - Image with a LoRA (SDXL + pixel art). On the owner PC (RTX 4060 Ti 16 GB) one 1024×1024 image took 80 s. - Voice note to summary (Whisper, then the local model). - Text to video (Wan 2.2 5B). - The LoRA style template now uses the pinned Qwen3 0.6B + TL;DR LoRA from the Models tab, so it works without your own files.
Not yet available:
- Text to speech: the Kokoro voices ship as .tar.bz2, which the wallet cannot unpack yet.
- Music (acestep.cpp): it has no versioned Windows release to pin; it will be built from source.
- Paid remote image and video jobs: these are the next phase.
Node categories¶
| Category | Nodes |
|---|---|
| AI Models | Select Model, Load Model, Inference; new in 0.8.9: Summarize, Classify, Extract Fields, Translate |
| Agents | AI Agent (read-only tools, step limit), Tool, Memory, Multi-Agent |
| Training | Dataset, Fine-Tuning, Evaluation |
| MeshLLM | Discover Provider, Route Request, Relay (AI P2P overlay) |
| Payments | Estimate Cost, Escrow, Settlement |
| Blockchain | Read Contract, Verify Transaction, Submit Transaction; new in 0.8.9: Block Height, Balance |
| Storage | IPFS Upload, Retrieve, Verify CID |
| Logic | Condition, Loop (runs a subworkflow per item), Router, Merge, Text Input; new in 0.8.9: Check |
| Output | Text, JSON, File, Result Viewer |
| Text & Data (new in 0.8.9) | Template, Transform Text, Extract Pattern, Parse JSON, Get Field, Build JSON, List, Calculate, Convert Units, Hash |
Ports are colour-coded by value type, and the editor refuses a connection between types that do not match. Fine-Tuning is in the library, but chain 20777 has no training market yet. The node therefore stops with that message and pays nothing.
Behaviour of the nodes added in 0.8.9¶
- Classify asks the model for one of two or three labels. It picks the label the answer names first, as a whole word. The input text goes out on that label's port, or on Other, and Label always carries the chosen name.
- Extract Fields keeps only the field names you list. A field the model did not find is
null. When the answer is not JSON, the object carries_error. - Get Field follows paths such as
result.items[0].name. A missing ornullvalue sends nothing on Value andfalseon Found, so later nodes are skipped rather than fed an empty value. - Check passes its value on when the test holds. Otherwise the node fails with your message, and the nodes after it are skipped.
- Extract Pattern uses .NET regular expressions with a 250 ms time limit and returns at most 100 matches. List and Transform Text stop at 1,000 items, and results are capped at 64 KiB.
- Calculate uses decimal arithmetic (no binary rounding errors) and refuses division by zero and overflow.
Convert Units is exact to 1 wei. Hash gives SHA-256 or Keccak-256 as
0xhex.
Datasets, knowledge bases and LoRA (CR-0054)¶
These nodes are in wallet 0.8.10 and later.
| Node | What it does |
|---|---|
| Load Documents | Reads .txt, .md, .jsonl, .json or .csv from the workflow data folder or an IPFS CID. Each JSONL or CSV row becomes a document; text comes from text/content/body, or prompt plus response/expected. |
| Text Splitter | Splits on paragraph, line, sentence and word boundaries into chunks of up to N characters. Chunks overlap, so a fact on a boundary is still found. |
| Knowledge Base | Builds a keyword (BM25) index. If an embedding model is connected, it also builds vectors. It can save a .kb.json file to the output folder, which you can share or pin on IPFS. |
| Load Knowledge Base | Opens a saved .kb.json from the data folder or IPFS. |
| Retriever | Finds the best passages: keyword, semantic (needs the embedding model the base was built with) or hybrid (both, combined by rank). |
| Knowledge Q&A | Answers from the retrieved passages only, citing them as [1], [2]. It says so when the passages do not contain the answer. |
| Knowledge Search Tool | Lets an AI Agent search the base itself (CALL knowledge_search <query>). |
| Documents to Dataset | Turns prompt/response rows into an evaluation dataset for Evaluation. |
| Model File | Serves any GGUF from the data folder (a chat model or an embedding model such as Qwen3-Embedding-0.6B) with your installed MeshLLM, on this PC only. |
| LoRA Loader | Adds a LoRA adapter (GGUF, from the data folder or IPFS) with a strength from −4 to 4. Chain loaders to stack up to 8 adapters, as in ComfyUI. |
| Apply LoRA | Merges the stack into a base model and serves the result locally. It can re-quantize to Q8_0 or Q4_K_M. |
How it works and its limits:
- Semantic search needs an embedding model. The Bonsai chat model cannot produce embeddings (MeshLLM refuses: "endpoint requires Embedding"). Connect a Model File node with an embedding GGUF to the Embedding model ports. Without one, use keyword search.
- LoRA is merged, not loaded at run time. MeshLLM cannot load adapters, so the wallet computes
W' = W + strength · (alpha / rank) · B·Afor every adapted weight. It writes a new GGUF, copying all other tensors unchanged, and caches it by base model, adapters, strengths and output type. The official llama.cpp release has no export tool, so this runs in the wallet itself. Only re-quantizing uses the officialllama-quantize(llama.cpp b12237, verified by SHA-256). The wallet asks before downloading it, and output F16 skips it. - Base models: F32, F16, BF16 or Q8_0 GGUF of the same family the LoRA was trained for. Bonsai 2 (ternary) cannot take a LoRA: its rotated ternary weights do not match standard adapters.
- Everything runs on your PC. A knowledge base, a LoRA or a merged model leaves it only when you upload it with an IPFS node.
Payments and signatures¶
- Every node that pays or signs (Escrow, Settlement, Submit Transaction, Fine-Tuning) is marked PAID or SIGNS. Before it runs, the wallet shows the maximum cost and asks you to approve.
- After your approval, the wallet's usual transaction dialog asks for your password. The workflow engine never holds a key or a password.
- Estimate Cost refuses a quote above the limit you set, before anything is signed.
- Escrow uses the AI market on chain 20777 (AIJobMarket
0xd5ACf35F5835e55B8e4328a51720D9F27836C8e4). The wallet checks the provider's on-chain charge against the answer it received, and refunds the unused budget.
Files: pqwerai-workflow/1¶
Export saves a workflow as a .pqflow.json file, and Import opens one. A workflow file holds data only:
- nodes, connections, groups and subworkflows;
- only node types from the catalog, and only the parameters each node declares;
- no file paths: Dataset and File nodes take a plain file name inside the wallet's workflow
dataandoutputfolders.
On import, the wallet refuses a file that breaks any of these rules or is larger than 1 MiB. Every Save version keeps a numbered copy, and every change is autosaved.
Editing¶
- Moving around: drag the canvas to pan, Ctrl+wheel to zoom, Fit to view, and a minimap.
- Adding and connecting: drag nodes from the library, or right-click the canvas and search. Drag from an output to an input to connect.
- Selecting: Ctrl+click selects several nodes, and Shift+drag selects with a box.
- Shortcuts: Ctrl+Z/Y undo and redo, Ctrl+C/V/D copy, paste and duplicate, Ctrl+G group, Ctrl+L auto-layout, Del deletes, F5 runs and Esc stops a run.
- Run status: each node shows its own state: waiting, running (with progress), waiting for approval, done, skipped, failed (with a fix) or cancelled.
Test status¶
Author review only. The tests ran on 15 November 2026:
| Test | What was checked |
|---|---|
| Engine unit tests (30; 70 with the 0.8.9 nodes) | Import rules, typed connections, cycles, branches, approval refusal, the cost limit, cancellation, loops, agent tools, CID checks, undo/redo, versions and the folder sandbox |
| Real-service test | The approved Bonsai model on MeshLLM and the public chain-20777 RPC: chat with memory, three-agent run, evaluation, agent with the chain-height tool, contract read, cancellation while the model answers, and a refused payment that never reaches signing |
| UI test | The real wallet app on chain 20777: templates, contract read, readable failures (no IPFS node, no local model, no provider), Advanced editing, problems list and undo |
Details: audit/evidence/2026-11-15-ai-workflows/ in the repository.
Sources at 7a33c77104: apps/pqwerai-wallet-windows/README.md · audit/evidence/2026-11-15-ai-workflows/README.md
Found a problem? Report a documentation issue · Build testnet-rc1+7a33c77104